フリー問題

NVIDIA-Certified Professional: Generative AI LLMs のフリー問題 1 / 20 問目

問題文

One deployment must serve three features from the same weights: a data extraction call that must be reproducible, a drafting call that should vary, and an audited summarization call whose output must be explainable later. Which design meets all three?

選択肢

  1. Pin one set of decoding settings for the whole deployment and choose the values that suit summarization; settings that differ between requests make the service's behavior impossible to reason about, and the extraction call can be made reproducible afterwards by comparing repeated calls instead.
  2. Use sampling for all three and raise the number of candidates for extraction, so that the most likely of the sampled candidates is the one returned.
  3. Run three separate deployments, one per feature, each with its own fixed settings, so that no request ever has to carry decoding settings of its own.
  4. Accept the decoding settings per request, default them to deterministic decoding, and record the settings actually used with each response.

解答・解説を確認するには

正解と解説の確認、回答の記録には無料登録が必要です。登録すると演習モードでフリー問題に回答し、正誤と解説をその場で確認できます。